xref: /freebsd/sys/vm/vm_pageout.c (revision a220d00e74dd245b4fca59c5eca0c53963686325)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
41  *
42  *
43  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44  * All rights reserved.
45  *
46  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  *
68  * $FreeBSD$
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include "opt_vm.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/lock.h>
80 #include <sys/mutex.h>
81 #include <sys/proc.h>
82 #include <sys/kthread.h>
83 #include <sys/ktr.h>
84 #include <sys/resourcevar.h>
85 #include <sys/signalvar.h>
86 #include <sys/vnode.h>
87 #include <sys/vmmeter.h>
88 #include <sys/sx.h>
89 #include <sys/sysctl.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_param.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vm_zone.h>
99 #include <vm/swap_pager.h>
100 #include <vm/vm_extern.h>
101 
102 #include <machine/mutex.h>
103 
104 /*
105  * System initialization
106  */
107 
108 /* the kernel process "vm_pageout"*/
109 static void vm_pageout __P((void));
110 static int vm_pageout_clean __P((vm_page_t));
111 static void vm_pageout_scan __P((int pass));
112 static int vm_pageout_free_page_calc __P((vm_size_t count));
113 struct proc *pageproc;
114 
115 static struct kproc_desc page_kp = {
116 	"pagedaemon",
117 	vm_pageout,
118 	&pageproc
119 };
120 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
121 
122 #if !defined(NO_SWAPPING)
123 /* the kernel process "vm_daemon"*/
124 static void vm_daemon __P((void));
125 static struct	proc *vmproc;
126 
127 static struct kproc_desc vm_kp = {
128 	"vmdaemon",
129 	vm_daemon,
130 	&vmproc
131 };
132 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
133 #endif
134 
135 
136 int vm_pages_needed=0;		/* Event on which pageout daemon sleeps */
137 int vm_pageout_deficit=0;	/* Estimated number of pages deficit */
138 int vm_pageout_pages_needed=0;	/* flag saying that the pageout daemon needs pages */
139 
140 #if !defined(NO_SWAPPING)
141 static int vm_pageout_req_swapout;	/* XXX */
142 static int vm_daemon_needed;
143 #endif
144 extern int vm_swap_size;
145 static int vm_max_launder = 32;
146 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
147 static int vm_pageout_full_stats_interval = 0;
148 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
149 static int defer_swap_pageouts=0;
150 static int disable_swap_pageouts=0;
151 
152 #if defined(NO_SWAPPING)
153 static int vm_swap_enabled=0;
154 static int vm_swap_idle_enabled=0;
155 #else
156 static int vm_swap_enabled=1;
157 static int vm_swap_idle_enabled=0;
158 #endif
159 
160 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
161 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
162 
163 SYSCTL_INT(_vm, OID_AUTO, max_launder,
164 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
165 
166 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
167 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
168 
169 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
170 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
171 
172 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
173 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
174 
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
176 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
177 
178 #if defined(NO_SWAPPING)
179 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
180 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
181 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
182 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
183 #else
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 #endif
189 
190 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
191 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
192 
193 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
194 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
195 
196 #define VM_PAGEOUT_PAGE_COUNT 16
197 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
198 
199 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
200 
201 #if !defined(NO_SWAPPING)
202 typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int));
203 static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t));
204 static freeer_fcn_t vm_pageout_object_deactivate_pages;
205 static void vm_req_vmdaemon __P((void));
206 #endif
207 static void vm_pageout_page_stats(void);
208 
209 /*
210  * vm_pageout_clean:
211  *
212  * Clean the page and remove it from the laundry.
213  *
214  * We set the busy bit to cause potential page faults on this page to
215  * block.  Note the careful timing, however, the busy bit isn't set till
216  * late and we cannot do anything that will mess with the page.
217  */
218 
219 static int
220 vm_pageout_clean(m)
221 	vm_page_t m;
222 {
223 	vm_object_t object;
224 	vm_page_t mc[2*vm_pageout_page_count];
225 	int pageout_count;
226 	int ib, is, page_base;
227 	vm_pindex_t pindex = m->pindex;
228 
229 	GIANT_REQUIRED;
230 
231 	object = m->object;
232 
233 	/*
234 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
235 	 * with the new swapper, but we could have serious problems paging
236 	 * out other object types if there is insufficient memory.
237 	 *
238 	 * Unfortunately, checking free memory here is far too late, so the
239 	 * check has been moved up a procedural level.
240 	 */
241 
242 	/*
243 	 * Don't mess with the page if it's busy, held, or special
244 	 */
245 	if ((m->hold_count != 0) ||
246 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
247 		return 0;
248 	}
249 
250 	mc[vm_pageout_page_count] = m;
251 	pageout_count = 1;
252 	page_base = vm_pageout_page_count;
253 	ib = 1;
254 	is = 1;
255 
256 	/*
257 	 * Scan object for clusterable pages.
258 	 *
259 	 * We can cluster ONLY if: ->> the page is NOT
260 	 * clean, wired, busy, held, or mapped into a
261 	 * buffer, and one of the following:
262 	 * 1) The page is inactive, or a seldom used
263 	 *    active page.
264 	 * -or-
265 	 * 2) we force the issue.
266 	 *
267 	 * During heavy mmap/modification loads the pageout
268 	 * daemon can really fragment the underlying file
269 	 * due to flushing pages out of order and not trying
270 	 * align the clusters (which leave sporatic out-of-order
271 	 * holes).  To solve this problem we do the reverse scan
272 	 * first and attempt to align our cluster, then do a
273 	 * forward scan if room remains.
274 	 */
275 
276 more:
277 	while (ib && pageout_count < vm_pageout_page_count) {
278 		vm_page_t p;
279 
280 		if (ib > pindex) {
281 			ib = 0;
282 			break;
283 		}
284 
285 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
286 			ib = 0;
287 			break;
288 		}
289 		if (((p->queue - p->pc) == PQ_CACHE) ||
290 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
291 			ib = 0;
292 			break;
293 		}
294 		vm_page_test_dirty(p);
295 		if ((p->dirty & p->valid) == 0 ||
296 		    p->queue != PQ_INACTIVE ||
297 		    p->wire_count != 0 ||	/* may be held by buf cache */
298 		    p->hold_count != 0) {	/* may be undergoing I/O */
299 			ib = 0;
300 			break;
301 		}
302 		mc[--page_base] = p;
303 		++pageout_count;
304 		++ib;
305 		/*
306 		 * alignment boundry, stop here and switch directions.  Do
307 		 * not clear ib.
308 		 */
309 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
310 			break;
311 	}
312 
313 	while (pageout_count < vm_pageout_page_count &&
314 	    pindex + is < object->size) {
315 		vm_page_t p;
316 
317 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
318 			break;
319 		if (((p->queue - p->pc) == PQ_CACHE) ||
320 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
321 			break;
322 		}
323 		vm_page_test_dirty(p);
324 		if ((p->dirty & p->valid) == 0 ||
325 		    p->queue != PQ_INACTIVE ||
326 		    p->wire_count != 0 ||	/* may be held by buf cache */
327 		    p->hold_count != 0) {	/* may be undergoing I/O */
328 			break;
329 		}
330 		mc[page_base + pageout_count] = p;
331 		++pageout_count;
332 		++is;
333 	}
334 
335 	/*
336 	 * If we exhausted our forward scan, continue with the reverse scan
337 	 * when possible, even past a page boundry.  This catches boundry
338 	 * conditions.
339 	 */
340 	if (ib && pageout_count < vm_pageout_page_count)
341 		goto more;
342 
343 	/*
344 	 * we allow reads during pageouts...
345 	 */
346 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
347 }
348 
349 /*
350  * vm_pageout_flush() - launder the given pages
351  *
352  *	The given pages are laundered.  Note that we setup for the start of
353  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
354  *	reference count all in here rather then in the parent.  If we want
355  *	the parent to do more sophisticated things we may have to change
356  *	the ordering.
357  */
358 
359 int
360 vm_pageout_flush(mc, count, flags)
361 	vm_page_t *mc;
362 	int count;
363 	int flags;
364 {
365 	vm_object_t object;
366 	int pageout_status[count];
367 	int numpagedout = 0;
368 	int i;
369 
370 	GIANT_REQUIRED;
371 	/*
372 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
373 	 * mark the pages read-only.
374 	 *
375 	 * We do not have to fixup the clean/dirty bits here... we can
376 	 * allow the pager to do it after the I/O completes.
377 	 *
378 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
379 	 * edge case with file fragments.
380 	 */
381 
382 	for (i = 0; i < count; i++) {
383 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
384 		vm_page_io_start(mc[i]);
385 		vm_page_protect(mc[i], VM_PROT_READ);
386 	}
387 
388 	object = mc[0]->object;
389 	vm_object_pip_add(object, count);
390 
391 	vm_pager_put_pages(object, mc, count,
392 	    (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)),
393 	    pageout_status);
394 
395 	for (i = 0; i < count; i++) {
396 		vm_page_t mt = mc[i];
397 
398 		switch (pageout_status[i]) {
399 		case VM_PAGER_OK:
400 			numpagedout++;
401 			break;
402 		case VM_PAGER_PEND:
403 			numpagedout++;
404 			break;
405 		case VM_PAGER_BAD:
406 			/*
407 			 * Page outside of range of object. Right now we
408 			 * essentially lose the changes by pretending it
409 			 * worked.
410 			 */
411 			pmap_clear_modify(mt);
412 			vm_page_undirty(mt);
413 			break;
414 		case VM_PAGER_ERROR:
415 		case VM_PAGER_FAIL:
416 			/*
417 			 * If page couldn't be paged out, then reactivate the
418 			 * page so it doesn't clog the inactive list.  (We
419 			 * will try paging out it again later).
420 			 */
421 			vm_page_activate(mt);
422 			break;
423 		case VM_PAGER_AGAIN:
424 			break;
425 		}
426 
427 		/*
428 		 * If the operation is still going, leave the page busy to
429 		 * block all other accesses. Also, leave the paging in
430 		 * progress indicator set so that we don't attempt an object
431 		 * collapse.
432 		 */
433 		if (pageout_status[i] != VM_PAGER_PEND) {
434 			vm_object_pip_wakeup(object);
435 			vm_page_io_finish(mt);
436 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
437 				vm_page_protect(mt, VM_PROT_READ);
438 		}
439 	}
440 	return numpagedout;
441 }
442 
443 #if !defined(NO_SWAPPING)
444 /*
445  *	vm_pageout_object_deactivate_pages
446  *
447  *	deactivate enough pages to satisfy the inactive target
448  *	requirements or if vm_page_proc_limit is set, then
449  *	deactivate all of the pages in the object and its
450  *	backing_objects.
451  *
452  *	The object and map must be locked.
453  */
454 static void
455 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only)
456 	vm_map_t map;
457 	vm_object_t object;
458 	vm_pindex_t desired;
459 	int map_remove_only;
460 {
461 	vm_page_t p, next;
462 	int rcount;
463 	int remove_mode;
464 
465 	GIANT_REQUIRED;
466 	if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
467 		return;
468 
469 	while (object) {
470 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
471 			return;
472 		if (object->paging_in_progress)
473 			return;
474 
475 		remove_mode = map_remove_only;
476 		if (object->shadow_count > 1)
477 			remove_mode = 1;
478 	/*
479 	 * scan the objects entire memory queue
480 	 */
481 		rcount = object->resident_page_count;
482 		p = TAILQ_FIRST(&object->memq);
483 		while (p && (rcount-- > 0)) {
484 			int actcount;
485 			if (pmap_resident_count(vm_map_pmap(map)) <= desired)
486 				return;
487 			next = TAILQ_NEXT(p, listq);
488 			cnt.v_pdpages++;
489 			if (p->wire_count != 0 ||
490 			    p->hold_count != 0 ||
491 			    p->busy != 0 ||
492 			    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
493 			    !pmap_page_exists(vm_map_pmap(map), p)) {
494 				p = next;
495 				continue;
496 			}
497 
498 			actcount = pmap_ts_referenced(p);
499 			if (actcount) {
500 				vm_page_flag_set(p, PG_REFERENCED);
501 			} else if (p->flags & PG_REFERENCED) {
502 				actcount = 1;
503 			}
504 
505 			if ((p->queue != PQ_ACTIVE) &&
506 				(p->flags & PG_REFERENCED)) {
507 				vm_page_activate(p);
508 				p->act_count += actcount;
509 				vm_page_flag_clear(p, PG_REFERENCED);
510 			} else if (p->queue == PQ_ACTIVE) {
511 				if ((p->flags & PG_REFERENCED) == 0) {
512 					p->act_count -= min(p->act_count, ACT_DECLINE);
513 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
514 						vm_page_protect(p, VM_PROT_NONE);
515 						vm_page_deactivate(p);
516 					} else {
517 						vm_pageq_requeue(p);
518 					}
519 				} else {
520 					vm_page_activate(p);
521 					vm_page_flag_clear(p, PG_REFERENCED);
522 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
523 						p->act_count += ACT_ADVANCE;
524 					vm_pageq_requeue(p);
525 				}
526 			} else if (p->queue == PQ_INACTIVE) {
527 				vm_page_protect(p, VM_PROT_NONE);
528 			}
529 			p = next;
530 		}
531 		object = object->backing_object;
532 	}
533 	return;
534 }
535 
536 /*
537  * deactivate some number of pages in a map, try to do it fairly, but
538  * that is really hard to do.
539  */
540 static void
541 vm_pageout_map_deactivate_pages(map, desired)
542 	vm_map_t map;
543 	vm_pindex_t desired;
544 {
545 	vm_map_entry_t tmpe;
546 	vm_object_t obj, bigobj;
547 	int nothingwired;
548 
549 	GIANT_REQUIRED;
550 	if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curthread)) {
551 		return;
552 	}
553 
554 	bigobj = NULL;
555 	nothingwired = TRUE;
556 
557 	/*
558 	 * first, search out the biggest object, and try to free pages from
559 	 * that.
560 	 */
561 	tmpe = map->header.next;
562 	while (tmpe != &map->header) {
563 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
564 			obj = tmpe->object.vm_object;
565 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
566 				((bigobj == NULL) ||
567 				 (bigobj->resident_page_count < obj->resident_page_count))) {
568 				bigobj = obj;
569 			}
570 		}
571 		if (tmpe->wired_count > 0)
572 			nothingwired = FALSE;
573 		tmpe = tmpe->next;
574 	}
575 
576 	if (bigobj)
577 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
578 
579 	/*
580 	 * Next, hunt around for other pages to deactivate.  We actually
581 	 * do this search sort of wrong -- .text first is not the best idea.
582 	 */
583 	tmpe = map->header.next;
584 	while (tmpe != &map->header) {
585 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
586 			break;
587 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
588 			obj = tmpe->object.vm_object;
589 			if (obj)
590 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
591 		}
592 		tmpe = tmpe->next;
593 	};
594 
595 	/*
596 	 * Remove all mappings if a process is swapped out, this will free page
597 	 * table pages.
598 	 */
599 	if (desired == 0 && nothingwired)
600 		pmap_remove(vm_map_pmap(map),
601 			VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
602 	vm_map_unlock(map);
603 	return;
604 }
605 #endif
606 
607 /*
608  * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore
609  * to vnode deadlocks.  We only do it for OBJT_DEFAULT and OBJT_SWAP objects
610  * which we know can be trivially freed.
611  */
612 
613 void
614 vm_pageout_page_free(vm_page_t m) {
615 	vm_object_t object = m->object;
616 	int type = object->type;
617 
618 	GIANT_REQUIRED;
619 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
620 		vm_object_reference(object);
621 	vm_page_busy(m);
622 	vm_page_protect(m, VM_PROT_NONE);
623 	vm_page_free(m);
624 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
625 		vm_object_deallocate(object);
626 }
627 
628 /*
629  *	vm_pageout_scan does the dirty work for the pageout daemon.
630  */
631 static void
632 vm_pageout_scan(int pass)
633 {
634 	vm_page_t m, next;
635 	struct vm_page marker;
636 	int save_page_shortage;
637 	int save_inactive_count;
638 	int page_shortage, maxscan, pcount;
639 	int addl_page_shortage, addl_page_shortage_init;
640 	struct proc *p, *bigproc;
641 	vm_offset_t size, bigsize;
642 	vm_object_t object;
643 	int actcount;
644 	int vnodes_skipped = 0;
645 	int maxlaunder;
646 	int s;
647 
648 	GIANT_REQUIRED;
649 	/*
650 	 * Do whatever cleanup that the pmap code can.
651 	 */
652 	pmap_collect();
653 
654 	addl_page_shortage_init = vm_pageout_deficit;
655 	vm_pageout_deficit = 0;
656 
657 	/*
658 	 * Calculate the number of pages we want to either free or move
659 	 * to the cache.
660 	 */
661 	page_shortage = vm_paging_target() + addl_page_shortage_init;
662 	save_page_shortage = page_shortage;
663 	save_inactive_count = cnt.v_inactive_count;
664 
665 	/*
666 	 * Initialize our marker
667 	 */
668 	bzero(&marker, sizeof(marker));
669 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
670 	marker.queue = PQ_INACTIVE;
671 	marker.wire_count = 1;
672 
673 	/*
674 	 * Start scanning the inactive queue for pages we can move to the
675 	 * cache or free.  The scan will stop when the target is reached or
676 	 * we have scanned the entire inactive queue.  Note that m->act_count
677 	 * is not used to form decisions for the inactive queue, only for the
678 	 * active queue.
679 	 *
680 	 * maxlaunder limits the number of dirty pages we flush per scan.
681 	 * For most systems a smaller value (16 or 32) is more robust under
682 	 * extreme memory and disk pressure because any unnecessary writes
683 	 * to disk can result in extreme performance degredation.  However,
684 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
685 	 * used) will die horribly with limited laundering.  If the pageout
686 	 * daemon cannot clean enough pages in the first pass, we let it go
687 	 * all out in succeeding passes.
688 	 */
689 
690 	if ((maxlaunder = vm_max_launder) <= 1)
691 		maxlaunder = 1;
692 	if (pass)
693 		maxlaunder = 10000;
694 
695 rescan0:
696 	addl_page_shortage = addl_page_shortage_init;
697 	maxscan = cnt.v_inactive_count;
698 
699 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
700 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
701 	     m = next) {
702 
703 		cnt.v_pdpages++;
704 
705 		if (m->queue != PQ_INACTIVE) {
706 			goto rescan0;
707 		}
708 
709 		next = TAILQ_NEXT(m, pageq);
710 
711 		/*
712 		 * skip marker pages
713 		 */
714 		if (m->flags & PG_MARKER)
715 			continue;
716 
717 		/*
718 		 * A held page may be undergoing I/O, so skip it.
719 		 */
720 		if (m->hold_count) {
721 			vm_pageq_requeue(m);
722 			addl_page_shortage++;
723 			continue;
724 		}
725 		/*
726 		 * Dont mess with busy pages, keep in the front of the
727 		 * queue, most likely are being paged out.
728 		 */
729 		if (m->busy || (m->flags & PG_BUSY)) {
730 			addl_page_shortage++;
731 			continue;
732 		}
733 
734 		/*
735 		 * If the object is not being used, we ignore previous
736 		 * references.
737 		 */
738 		if (m->object->ref_count == 0) {
739 			vm_page_flag_clear(m, PG_REFERENCED);
740 			pmap_clear_reference(m);
741 
742 		/*
743 		 * Otherwise, if the page has been referenced while in the
744 		 * inactive queue, we bump the "activation count" upwards,
745 		 * making it less likely that the page will be added back to
746 		 * the inactive queue prematurely again.  Here we check the
747 		 * page tables (or emulated bits, if any), given the upper
748 		 * level VM system not knowing anything about existing
749 		 * references.
750 		 */
751 		} else if (((m->flags & PG_REFERENCED) == 0) &&
752 			(actcount = pmap_ts_referenced(m))) {
753 			vm_page_activate(m);
754 			m->act_count += (actcount + ACT_ADVANCE);
755 			continue;
756 		}
757 
758 		/*
759 		 * If the upper level VM system knows about any page
760 		 * references, we activate the page.  We also set the
761 		 * "activation count" higher than normal so that we will less
762 		 * likely place pages back onto the inactive queue again.
763 		 */
764 		if ((m->flags & PG_REFERENCED) != 0) {
765 			vm_page_flag_clear(m, PG_REFERENCED);
766 			actcount = pmap_ts_referenced(m);
767 			vm_page_activate(m);
768 			m->act_count += (actcount + ACT_ADVANCE + 1);
769 			continue;
770 		}
771 
772 		/*
773 		 * If the upper level VM system doesn't know anything about
774 		 * the page being dirty, we have to check for it again.  As
775 		 * far as the VM code knows, any partially dirty pages are
776 		 * fully dirty.
777 		 */
778 		if (m->dirty == 0) {
779 			vm_page_test_dirty(m);
780 		} else {
781 			vm_page_dirty(m);
782 		}
783 
784 		/*
785 		 * Invalid pages can be easily freed
786 		 */
787 		if (m->valid == 0) {
788 			vm_pageout_page_free(m);
789 			cnt.v_dfree++;
790 			--page_shortage;
791 
792 		/*
793 		 * Clean pages can be placed onto the cache queue.  This
794 		 * effectively frees them.
795 		 */
796 		} else if (m->dirty == 0) {
797 			vm_page_cache(m);
798 			--page_shortage;
799 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
800 			/*
801 			 * Dirty pages need to be paged out, but flushing
802 			 * a page is extremely expensive verses freeing
803 			 * a clean page.  Rather then artificially limiting
804 			 * the number of pages we can flush, we instead give
805 			 * dirty pages extra priority on the inactive queue
806 			 * by forcing them to be cycled through the queue
807 			 * twice before being flushed, after which the
808 			 * (now clean) page will cycle through once more
809 			 * before being freed.  This significantly extends
810 			 * the thrash point for a heavily loaded machine.
811 			 */
812 			vm_page_flag_set(m, PG_WINATCFLS);
813 			vm_pageq_requeue(m);
814 		} else if (maxlaunder > 0) {
815 			/*
816 			 * We always want to try to flush some dirty pages if
817 			 * we encounter them, to keep the system stable.
818 			 * Normally this number is small, but under extreme
819 			 * pressure where there are insufficient clean pages
820 			 * on the inactive queue, we may have to go all out.
821 			 */
822 			int swap_pageouts_ok;
823 			struct vnode *vp = NULL;
824 			struct mount *mp;
825 
826 			object = m->object;
827 
828 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
829 				swap_pageouts_ok = 1;
830 			} else {
831 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
832 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
833 				vm_page_count_min());
834 
835 			}
836 
837 			/*
838 			 * We don't bother paging objects that are "dead".
839 			 * Those objects are in a "rundown" state.
840 			 */
841 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
842 				vm_pageq_requeue(m);
843 				continue;
844 			}
845 
846 			/*
847 			 * The object is already known NOT to be dead.   It
848 			 * is possible for the vget() to block the whole
849 			 * pageout daemon, but the new low-memory handling
850 			 * code should prevent it.
851 			 *
852 			 * The previous code skipped locked vnodes and, worse,
853 			 * reordered pages in the queue.  This results in
854 			 * completely non-deterministic operation and, on a
855 			 * busy system, can lead to extremely non-optimal
856 			 * pageouts.  For example, it can cause clean pages
857 			 * to be freed and dirty pages to be moved to the end
858 			 * of the queue.  Since dirty pages are also moved to
859 			 * the end of the queue once-cleaned, this gives
860 			 * way too large a weighting to defering the freeing
861 			 * of dirty pages.
862 			 *
863 			 * XXX we need to be able to apply a timeout to the
864 			 * vget() lock attempt.
865 			 */
866 
867 			if (object->type == OBJT_VNODE) {
868 				vp = object->handle;
869 
870 				mp = NULL;
871 				if (vp->v_type == VREG)
872 					vn_start_write(vp, &mp, V_NOWAIT);
873 				if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ, curthread)) {
874 					vn_finished_write(mp);
875 					if (object->flags & OBJ_MIGHTBEDIRTY)
876 						vnodes_skipped++;
877 					continue;
878 				}
879 
880 				/*
881 				 * The page might have been moved to another
882 				 * queue during potential blocking in vget()
883 				 * above.  The page might have been freed and
884 				 * reused for another vnode.  The object might
885 				 * have been reused for another vnode.
886 				 */
887 				if (m->queue != PQ_INACTIVE ||
888 				    m->object != object ||
889 				    object->handle != vp) {
890 					if (object->flags & OBJ_MIGHTBEDIRTY)
891 						vnodes_skipped++;
892 					vput(vp);
893 					vn_finished_write(mp);
894 					continue;
895 				}
896 
897 				/*
898 				 * The page may have been busied during the
899 				 * blocking in vput();  We don't move the
900 				 * page back onto the end of the queue so that
901 				 * statistics are more correct if we don't.
902 				 */
903 				if (m->busy || (m->flags & PG_BUSY)) {
904 					vput(vp);
905 					vn_finished_write(mp);
906 					continue;
907 				}
908 
909 				/*
910 				 * If the page has become held it might
911 				 * be undergoing I/O, so skip it
912 				 */
913 				if (m->hold_count) {
914 					vm_pageq_requeue(m);
915 					if (object->flags & OBJ_MIGHTBEDIRTY)
916 						vnodes_skipped++;
917 					vput(vp);
918 					vn_finished_write(mp);
919 					continue;
920 				}
921 			}
922 
923 			/*
924 			 * If a page is dirty, then it is either being washed
925 			 * (but not yet cleaned) or it is still in the
926 			 * laundry.  If it is still in the laundry, then we
927 			 * start the cleaning operation.
928 			 *
929 			 * This operation may cluster, invalidating the 'next'
930 			 * pointer.  To prevent an inordinate number of
931 			 * restarts we use our marker to remember our place.
932 			 *
933 			 * decrement page_shortage on success to account for
934 			 * the (future) cleaned page.  Otherwise we could wind
935 			 * up laundering or cleaning too many pages.
936 			 */
937 			s = splvm();
938 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
939 			splx(s);
940 			if (vm_pageout_clean(m) != 0) {
941 				--page_shortage;
942 				--maxlaunder;
943 			}
944 			s = splvm();
945 			next = TAILQ_NEXT(&marker, pageq);
946 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
947 			splx(s);
948 			if (vp) {
949 				vput(vp);
950 				vn_finished_write(mp);
951 			}
952 		}
953 	}
954 
955 	/*
956 	 * Compute the number of pages we want to try to move from the
957 	 * active queue to the inactive queue.
958 	 */
959 	page_shortage = vm_paging_target() +
960 		cnt.v_inactive_target - cnt.v_inactive_count;
961 	page_shortage += addl_page_shortage;
962 
963 	/*
964 	 * Scan the active queue for things we can deactivate. We nominally
965 	 * track the per-page activity counter and use it to locate
966 	 * deactivation candidates.
967 	 */
968 
969 	pcount = cnt.v_active_count;
970 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
971 
972 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
973 
974 		/*
975 		 * This is a consistency check, and should likely be a panic
976 		 * or warning.
977 		 */
978 		if (m->queue != PQ_ACTIVE) {
979 			break;
980 		}
981 
982 		next = TAILQ_NEXT(m, pageq);
983 		/*
984 		 * Don't deactivate pages that are busy.
985 		 */
986 		if ((m->busy != 0) ||
987 		    (m->flags & PG_BUSY) ||
988 		    (m->hold_count != 0)) {
989 			vm_pageq_requeue(m);
990 			m = next;
991 			continue;
992 		}
993 
994 		/*
995 		 * The count for pagedaemon pages is done after checking the
996 		 * page for eligibility...
997 		 */
998 		cnt.v_pdpages++;
999 
1000 		/*
1001 		 * Check to see "how much" the page has been used.
1002 		 */
1003 		actcount = 0;
1004 		if (m->object->ref_count != 0) {
1005 			if (m->flags & PG_REFERENCED) {
1006 				actcount += 1;
1007 			}
1008 			actcount += pmap_ts_referenced(m);
1009 			if (actcount) {
1010 				m->act_count += ACT_ADVANCE + actcount;
1011 				if (m->act_count > ACT_MAX)
1012 					m->act_count = ACT_MAX;
1013 			}
1014 		}
1015 
1016 		/*
1017 		 * Since we have "tested" this bit, we need to clear it now.
1018 		 */
1019 		vm_page_flag_clear(m, PG_REFERENCED);
1020 
1021 		/*
1022 		 * Only if an object is currently being used, do we use the
1023 		 * page activation count stats.
1024 		 */
1025 		if (actcount && (m->object->ref_count != 0)) {
1026 			vm_pageq_requeue(m);
1027 		} else {
1028 			m->act_count -= min(m->act_count, ACT_DECLINE);
1029 			if (vm_pageout_algorithm ||
1030 			    m->object->ref_count == 0 ||
1031 			    m->act_count == 0) {
1032 				page_shortage--;
1033 				if (m->object->ref_count == 0) {
1034 					vm_page_protect(m, VM_PROT_NONE);
1035 					if (m->dirty == 0)
1036 						vm_page_cache(m);
1037 					else
1038 						vm_page_deactivate(m);
1039 				} else {
1040 					vm_page_deactivate(m);
1041 				}
1042 			} else {
1043 				vm_pageq_requeue(m);
1044 			}
1045 		}
1046 		m = next;
1047 	}
1048 
1049 	s = splvm();
1050 
1051 	/*
1052 	 * We try to maintain some *really* free pages, this allows interrupt
1053 	 * code to be guaranteed space.  Since both cache and free queues
1054 	 * are considered basically 'free', moving pages from cache to free
1055 	 * does not effect other calculations.
1056 	 */
1057 
1058 	while (cnt.v_free_count < cnt.v_free_reserved) {
1059 		static int cache_rover = 0;
1060 		m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE);
1061 		if (!m)
1062 			break;
1063 		if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1064 		    m->busy ||
1065 		    m->hold_count ||
1066 		    m->wire_count) {
1067 #ifdef INVARIANTS
1068 			printf("Warning: busy page %p found in cache\n", m);
1069 #endif
1070 			vm_page_deactivate(m);
1071 			continue;
1072 		}
1073 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1074 		vm_pageout_page_free(m);
1075 		cnt.v_dfree++;
1076 	}
1077 	splx(s);
1078 
1079 #if !defined(NO_SWAPPING)
1080 	/*
1081 	 * Idle process swapout -- run once per second.
1082 	 */
1083 	if (vm_swap_idle_enabled) {
1084 		static long lsec;
1085 		if (time_second != lsec) {
1086 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1087 			vm_req_vmdaemon();
1088 			lsec = time_second;
1089 		}
1090 	}
1091 #endif
1092 
1093 	/*
1094 	 * If we didn't get enough free pages, and we have skipped a vnode
1095 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1096 	 * if we did not get enough free pages.
1097 	 */
1098 	if (vm_paging_target() > 0) {
1099 		if (vnodes_skipped && vm_page_count_min())
1100 			(void) speedup_syncer();
1101 #if !defined(NO_SWAPPING)
1102 		if (vm_swap_enabled && vm_page_count_target()) {
1103 			vm_req_vmdaemon();
1104 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1105 		}
1106 #endif
1107 	}
1108 
1109 	/*
1110 	 * If we are out of swap and were not able to reach our paging
1111 	 * target, kill the largest process.
1112 	 *
1113 	 * We keep the process bigproc locked once we find it to keep anyone
1114 	 * from messing with it; however, there is a possibility of
1115 	 * deadlock if process B is bigproc and one of it's child processes
1116 	 * attempts to propagate a signal to B while we are waiting for A's
1117 	 * lock while walking this list.  To avoid this, we don't block on
1118 	 * the process lock but just skip a process if it is already locked.
1119 	 */
1120 	if ((vm_swap_size < 64 && vm_page_count_min()) ||
1121 	    (swap_pager_full && vm_paging_target() > 0)) {
1122 #if 0
1123 	if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) {
1124 #endif
1125 		bigproc = NULL;
1126 		bigsize = 0;
1127 		sx_slock(&allproc_lock);
1128 		LIST_FOREACH(p, &allproc, p_list) {
1129 			/*
1130 			 * If this process is already locked, skip it.
1131 			 */
1132 			if (PROC_TRYLOCK(p) == 0)
1133 				continue;
1134 			/*
1135 			 * if this is a system process, skip it
1136 			 */
1137 			if ((p->p_flag & P_SYSTEM) || (p->p_lock > 0) ||
1138 			    (p->p_pid == 1) ||
1139 			    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1140 				PROC_UNLOCK(p);
1141 				continue;
1142 			}
1143 			/*
1144 			 * if the process is in a non-running type state,
1145 			 * don't touch it.
1146 			 */
1147 			mtx_lock_spin(&sched_lock);
1148 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1149 				mtx_unlock_spin(&sched_lock);
1150 				PROC_UNLOCK(p);
1151 				continue;
1152 			}
1153 			mtx_unlock_spin(&sched_lock);
1154 			/*
1155 			 * get the process size
1156 			 */
1157 			size = vmspace_resident_count(p->p_vmspace) +
1158 				vmspace_swap_count(p->p_vmspace);
1159 			/*
1160 			 * if the this process is bigger than the biggest one
1161 			 * remember it.
1162 			 */
1163 			if (size > bigsize) {
1164 				if (bigproc != NULL)
1165 					PROC_UNLOCK(bigproc);
1166 				bigproc = p;
1167 				bigsize = size;
1168 			} else
1169 				PROC_UNLOCK(p);
1170 		}
1171 		sx_sunlock(&allproc_lock);
1172 		if (bigproc != NULL) {
1173 			struct ksegrp *kg;
1174 			killproc(bigproc, "out of swap space");
1175 			mtx_lock_spin(&sched_lock);
1176 			FOREACH_KSEGRP_IN_PROC(bigproc, kg) {
1177 				kg->kg_estcpu = 0;
1178 				kg->kg_nice = PRIO_MIN; /* XXXKSE ??? */
1179 				resetpriority(kg);
1180 			}
1181 			mtx_unlock_spin(&sched_lock);
1182 			PROC_UNLOCK(bigproc);
1183 			wakeup(&cnt.v_free_count);
1184 		}
1185 	}
1186 }
1187 
1188 /*
1189  * This routine tries to maintain the pseudo LRU active queue,
1190  * so that during long periods of time where there is no paging,
1191  * that some statistic accumulation still occurs.  This code
1192  * helps the situation where paging just starts to occur.
1193  */
1194 static void
1195 vm_pageout_page_stats()
1196 {
1197 	vm_page_t m,next;
1198 	int pcount,tpcount;		/* Number of pages to check */
1199 	static int fullintervalcount = 0;
1200 	int page_shortage;
1201 	int s0;
1202 
1203 	page_shortage =
1204 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1205 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1206 
1207 	if (page_shortage <= 0)
1208 		return;
1209 
1210 	s0 = splvm();
1211 
1212 	pcount = cnt.v_active_count;
1213 	fullintervalcount += vm_pageout_stats_interval;
1214 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1215 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1216 		if (pcount > tpcount)
1217 			pcount = tpcount;
1218 	} else {
1219 		fullintervalcount = 0;
1220 	}
1221 
1222 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1223 	while ((m != NULL) && (pcount-- > 0)) {
1224 		int actcount;
1225 
1226 		if (m->queue != PQ_ACTIVE) {
1227 			break;
1228 		}
1229 
1230 		next = TAILQ_NEXT(m, pageq);
1231 		/*
1232 		 * Don't deactivate pages that are busy.
1233 		 */
1234 		if ((m->busy != 0) ||
1235 		    (m->flags & PG_BUSY) ||
1236 		    (m->hold_count != 0)) {
1237 			vm_pageq_requeue(m);
1238 			m = next;
1239 			continue;
1240 		}
1241 
1242 		actcount = 0;
1243 		if (m->flags & PG_REFERENCED) {
1244 			vm_page_flag_clear(m, PG_REFERENCED);
1245 			actcount += 1;
1246 		}
1247 
1248 		actcount += pmap_ts_referenced(m);
1249 		if (actcount) {
1250 			m->act_count += ACT_ADVANCE + actcount;
1251 			if (m->act_count > ACT_MAX)
1252 				m->act_count = ACT_MAX;
1253 			vm_pageq_requeue(m);
1254 		} else {
1255 			if (m->act_count == 0) {
1256 				/*
1257 				 * We turn off page access, so that we have
1258 				 * more accurate RSS stats.  We don't do this
1259 				 * in the normal page deactivation when the
1260 				 * system is loaded VM wise, because the
1261 				 * cost of the large number of page protect
1262 				 * operations would be higher than the value
1263 				 * of doing the operation.
1264 				 */
1265 				vm_page_protect(m, VM_PROT_NONE);
1266 				vm_page_deactivate(m);
1267 			} else {
1268 				m->act_count -= min(m->act_count, ACT_DECLINE);
1269 				vm_pageq_requeue(m);
1270 			}
1271 		}
1272 
1273 		m = next;
1274 	}
1275 	splx(s0);
1276 }
1277 
1278 static int
1279 vm_pageout_free_page_calc(count)
1280 vm_size_t count;
1281 {
1282 	if (count < cnt.v_page_count)
1283 		 return 0;
1284 	/*
1285 	 * free_reserved needs to include enough for the largest swap pager
1286 	 * structures plus enough for any pv_entry structs when paging.
1287 	 */
1288 	if (cnt.v_page_count > 1024)
1289 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1290 	else
1291 		cnt.v_free_min = 4;
1292 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1293 		cnt.v_interrupt_free_min;
1294 	cnt.v_free_reserved = vm_pageout_page_count +
1295 		cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1296 	cnt.v_free_severe = cnt.v_free_min / 2;
1297 	cnt.v_free_min += cnt.v_free_reserved;
1298 	cnt.v_free_severe += cnt.v_free_reserved;
1299 	return 1;
1300 }
1301 
1302 
1303 /*
1304  *	vm_pageout is the high level pageout daemon.
1305  */
1306 static void
1307 vm_pageout()
1308 {
1309 	int pass;
1310 
1311 	mtx_lock(&Giant);
1312 
1313 	/*
1314 	 * Initialize some paging parameters.
1315 	 */
1316 
1317 	cnt.v_interrupt_free_min = 2;
1318 	if (cnt.v_page_count < 2000)
1319 		vm_pageout_page_count = 8;
1320 
1321 	vm_pageout_free_page_calc(cnt.v_page_count);
1322 	/*
1323 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1324 	 * that these are more a measure of the VM cache queue hysteresis
1325 	 * then the VM free queue.  Specifically, v_free_target is the
1326 	 * high water mark (free+cache pages).
1327 	 *
1328 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1329 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1330 	 * be big enough to handle memory needs while the pageout daemon
1331 	 * is signalled and run to free more pages.
1332 	 */
1333 	if (cnt.v_free_count > 6144)
1334 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1335 	else
1336 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1337 
1338 	if (cnt.v_free_count > 2048) {
1339 		cnt.v_cache_min = cnt.v_free_target;
1340 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1341 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1342 	} else {
1343 		cnt.v_cache_min = 0;
1344 		cnt.v_cache_max = 0;
1345 		cnt.v_inactive_target = cnt.v_free_count / 4;
1346 	}
1347 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1348 		cnt.v_inactive_target = cnt.v_free_count / 3;
1349 
1350 	/* XXX does not really belong here */
1351 	if (vm_page_max_wired == 0)
1352 		vm_page_max_wired = cnt.v_free_count / 3;
1353 
1354 	if (vm_pageout_stats_max == 0)
1355 		vm_pageout_stats_max = cnt.v_free_target;
1356 
1357 	/*
1358 	 * Set interval in seconds for stats scan.
1359 	 */
1360 	if (vm_pageout_stats_interval == 0)
1361 		vm_pageout_stats_interval = 5;
1362 	if (vm_pageout_full_stats_interval == 0)
1363 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1364 
1365 
1366 	/*
1367 	 * Set maximum free per pass
1368 	 */
1369 	if (vm_pageout_stats_free_max == 0)
1370 		vm_pageout_stats_free_max = 5;
1371 
1372 	PROC_LOCK(curthread->td_proc);
1373 	curthread->td_proc->p_flag |= P_BUFEXHAUST;
1374 	PROC_UNLOCK(curthread->td_proc);
1375 	swap_pager_swap_init();
1376 	pass = 0;
1377 	/*
1378 	 * The pageout daemon is never done, so loop forever.
1379 	 */
1380 	while (TRUE) {
1381 		int error;
1382 		int s = splvm();
1383 
1384 		/*
1385 		 * If we have enough free memory, wakeup waiters.  Do
1386 		 * not clear vm_pages_needed until we reach our target,
1387 		 * otherwise we may be woken up over and over again and
1388 		 * waste a lot of cpu.
1389 		 */
1390 		if (vm_pages_needed && !vm_page_count_min()) {
1391 			if (vm_paging_needed() <= 0)
1392 				vm_pages_needed = 0;
1393 			wakeup(&cnt.v_free_count);
1394 		}
1395 		if (vm_pages_needed) {
1396 			/*
1397 			 * Still not done, take a second pass without waiting
1398 			 * (unlimited dirty cleaning), otherwise sleep a bit
1399 			 * and try again.
1400 			 */
1401 			++pass;
1402 			if (pass > 1)
1403 				tsleep(&vm_pages_needed, PVM,
1404 				       "psleep", hz/2);
1405 		} else {
1406 			/*
1407 			 * Good enough, sleep & handle stats.  Prime the pass
1408 			 * for the next run.
1409 			 */
1410 			if (pass > 1)
1411 				pass = 1;
1412 			else
1413 				pass = 0;
1414 			error = tsleep(&vm_pages_needed, PVM,
1415 				    "psleep", vm_pageout_stats_interval * hz);
1416 			if (error && !vm_pages_needed) {
1417 				splx(s);
1418 				pass = 0;
1419 				vm_pageout_page_stats();
1420 				continue;
1421 			}
1422 		}
1423 
1424 		if (vm_pages_needed)
1425 			cnt.v_pdwakeups++;
1426 		splx(s);
1427 		vm_pageout_scan(pass);
1428 		vm_pageout_deficit = 0;
1429 	}
1430 }
1431 
1432 void
1433 pagedaemon_wakeup()
1434 {
1435 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1436 		vm_pages_needed++;
1437 		wakeup(&vm_pages_needed);
1438 	}
1439 }
1440 
1441 #if !defined(NO_SWAPPING)
1442 static void
1443 vm_req_vmdaemon()
1444 {
1445 	static int lastrun = 0;
1446 
1447 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1448 		wakeup(&vm_daemon_needed);
1449 		lastrun = ticks;
1450 	}
1451 }
1452 
1453 static void
1454 vm_daemon()
1455 {
1456 	struct proc *p;
1457 
1458 	mtx_lock(&Giant);
1459 	while (TRUE) {
1460 		tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1461 		if (vm_pageout_req_swapout) {
1462 			swapout_procs(vm_pageout_req_swapout);
1463 			vm_pageout_req_swapout = 0;
1464 		}
1465 		/*
1466 		 * scan the processes for exceeding their rlimits or if
1467 		 * process is swapped out -- deactivate pages
1468 		 */
1469 
1470 		sx_slock(&allproc_lock);
1471 		LIST_FOREACH(p, &allproc, p_list) {
1472 			vm_pindex_t limit, size;
1473 
1474 			/*
1475 			 * if this is a system process or if we have already
1476 			 * looked at this process, skip it.
1477 			 */
1478 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1479 				continue;
1480 			}
1481 			/*
1482 			 * if the process is in a non-running type state,
1483 			 * don't touch it.
1484 			 */
1485 			mtx_lock_spin(&sched_lock);
1486 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1487 				mtx_unlock_spin(&sched_lock);
1488 				continue;
1489 			}
1490 			/*
1491 			 * get a limit
1492 			 */
1493 			limit = OFF_TO_IDX(
1494 			    qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1495 				p->p_rlimit[RLIMIT_RSS].rlim_max));
1496 
1497 			/*
1498 			 * let processes that are swapped out really be
1499 			 * swapped out set the limit to nothing (will force a
1500 			 * swap-out.)
1501 			 */
1502 			if ((p->p_sflag & PS_INMEM) == 0)
1503 				limit = 0;	/* XXX */
1504 			mtx_unlock_spin(&sched_lock);
1505 
1506 			size = vmspace_resident_count(p->p_vmspace);
1507 			if (limit >= 0 && size >= limit) {
1508 				vm_pageout_map_deactivate_pages(
1509 				    &p->p_vmspace->vm_map, limit);
1510 			}
1511 		}
1512 		sx_sunlock(&allproc_lock);
1513 	}
1514 }
1515 #endif
1516